Research On Fault Diagnosis Method Of Planetary Gearboxs By Multi-Channel Attention Mechanism

Wind turbines operate in complex and variable work environments,and transmission components such as gearboxes often generate fault under alternating loads causing unit shutdowns and seriously affecting equipment efficiency.The components of a planetary gearbox are numerous and interact with each other,and the raw data of the transmission components is coupled with each other during operation,resulting in strong noise signals submerging weak fault information,making it difficult to accurately identify and extract fault features and achieve real-time fault diagnosis when the devices are in running state.The intelligent fault diagnosis methods are based on convolutional neural networks,and the signals of each channel are mixed and identified through multi-channel convolutional layers.It can effectively extract the fault features of the hidden layers and complete high-precision fault category determination.Therefore,this paper utilizes the suffix tree method to encode the repetitive features of the vibration signals of the wind turbine planetary gearbox,converting the physical features into visual images,and using the attention mechanism to identify the fault information in the coding results in advance.Lastly,an adaptive convolutional neural network is introduced to complete the information fusion of the calibrated features,implementing supervising train of network models and completing high-precision model for intelligent fault diagnosis of planetary gearbox.This article simulates the operation process of the planetary gearbox of a wind turbine using the Spectra Quest comprehensive experimental platform.The main research work are as follows:Firstly,based on the characteristics of structural and gear transmission of the wind turbine planetary gearbox,and analyzed the common fault forms of the wind turbine and the forming mechanization of the fault parts,the theoretical impact frequency of the fault gear in the meshing process is derived,and based on the gear transmission characteristics,the suffix tree method is used to complete the repetitive feature coding of the vibration signal.This method directly completes data processing relying on time-domain signals with high signal integrity,strong timeliness,and intuitive image results,laying a good foundation for subsequent intelligent classification.Secondly,introduced the attention mechanism into the network model,using twodimensional visualization images encoded by repetitive features to obtain the feature information of multi-channel raw data.The identification effect of attention mechanism is analyzed based on theoretical collided frequency,and physical characteristics such as fault collided point location,repetitive feature length,and fault frequency are identified by combining the visualization images of repetitive features.Finally,the raw data volume of all fault categories after repeated feature coding is counted.After randomly disordering the sample order and dividing the training set and test,the network model architecture is determined.The parameters of the convolutional neural network are trained iteratively,and then,the data of the test set is imported to verify the classification results.The overall accuracy is up to 98.9%,and the Kappa coefficient is 0.986.To thoroughly illustrate the data transmission rule of convolutional neural networks,the identification effect of convolutional layers is analyzed through category activation mapping.The experimental results fully demonstrate that the method described in this article has significant improved in the identification accuracy of fault categories and reduced the time of the diagnosis process,which has important reference value and practical guidance significance and provides new ideas for the development of intelligent fault diagnosis field.